In a recent review published in the Life Journal, a group of authors reviewed image-processing techniques in machine learning (ML) for skin cancer detection using clinical images, evaluating their efficacy, available datasets, and challenges.
Study: Automatic Skin Cancer Detection Using Clinical Images: A Comprehensive Review. Image Credit: PopTika/Shutterstock.com
Background
Over recent decades, the incidence of skin cancer has surged, with melanoma diagnoses growing by 31% between 2012-2022 and accounting for 80% of related deaths. However, early detection can lead to survival rates as high as 99%, which drops to 30% if the disease metastasizes.
Pigmented skin lesions (PSLs), ranging from benign moles to malignant melanomas, are central to this concern. Dermoscopy, employing a magnifying lens, aids in diagnosing PSLs. Yet, non-specialists often rely on standard cameras, sending images for expert diagnosis—a method nearly as effective as in-person consultations.
Interestingly, dermatologists performed better with macroscopic than dermoscopic images in one study. Recently, computational methods, particularly ML, have shown promise in assisting early skin cancer diagnosis.
Further research is needed because early melanoma detection is crucial for effective treatment. Yet, current diagnostic methods may not be consistently accessible or reliable, highlighting the potential of computer-aided tools in bridging the detection gap.
The imperative of early melanoma detection
Melanoma, a highly fatal form of skin cancer, often gets detected in advanced stages. However, studies from the past decade have indicated that early diagnosis can drastically reduce mortality rates.
In this endeavor, many have utilized imaging methods and artificial intelligence to diagnose these malignancies. Various unique and conventional methodologies have been proposed to address these diagnostic challenges.
Existing literature and gaps
Although there is a plethora of evaluations on identifying skin cancer through artificial intelligence, a gap exists in the comprehensive analysis of diagnosing skin cancer using clinical images and machine learning.
Few have showcased a complete overview of all available clinical datasets. For clarity, the authors compared the review with other recent reviews, considering year scope, imaging modality type, and major tasks in the automated skin cancer detection pipeline.
Public datasets of clinical skin images
There are numerous public datasets with clinical skin images utilized by various teams over the past decade. Prominent among them are DermQuest and MED-NODE. However, different datasets have led to non-comparable error metrics in the reviewed papers.
Image preprocessing techniques
Images captured frequently contain artifacts, making segmentation a challenge. Preprocessing corrects these irregularities, like illumination anomalies or the presence of hairs, ensuring algorithms perform accurately. The majority of the reviewed literature employs four main types of preprocessing methods.
Illumination correction
Images might contain illumination artifacts. To prevent shading and lesion border confusion, shading is diminished before segmentation.
Various strategies have been used to eliminate such irregularities, such as the data-driven method in “Hue, Saturation, Value” (HSV) color space and other techniques like thresholding algorithms.
Artifact removal
Artifacts like noise, skin lines, or hair can affect the image quality. Various tools and methods have been developed to mitigate these effects, including DullRazor for hair removal and Gaussian filters for noise reduction. However, the efficacy of these methods varies and must be used judiciously to avoid undermining machine learning training.
Image resizing and cropping
Ensuring uniformity in image size is vital for training convolutional neural network (CNN) models. This is achieved through cropping, resizing, and re-scaling the images.
Data augmentation
Enhancing machine learning performance involves creating varied training examples. This is done through data augmentation, which is especially useful for imbalanced datasets. Techniques include cropping, rotation, and noise addition, among others.
Other preprocessing methods
Various researchers, such as contrast enhancement, histogram equalization, and the use of algorithms like FastCUT explored different approaches. These methods aim to improve the overall quality and reliability of the dataset.
Image segmentation in dermatology
Image segmentation involves partitioning an image into different sections or pixel clusters, called image objects. It simplifies image analysis, aiding in easier lesion extraction. However, skin image segmentation remains challenging, often necessitating pre and post-processing.
Predominant segmentation approaches
Numerous techniques for image segmentation have been explored. Two commonly used methods are Otsu's method and K-means clustering. Few researchers employed the standard Otsu method, while others combined Otsu's method with other techniques, achieving 100% accuracy in lesion extent determination.
However, such claims’ reliability remains debatable. In contrast, numerous researchers used the K-means clustering algorithm, with variances in accuracy due to diverse datasets.
Alternative segmentation methods
Other explored segmentation techniques include the Chan–Vese active contour method, fast independent component analysis, and synthesis and convergence of intermediate decaying omnigradients (SCIDOG).
These methods have reported various performance metrics, but the absence of consistent evaluation across methods limits performance comparisons.
Post-segmentation processing
After segmentation, post-processing techniques enhance the segmented image. Widely used methods include morphological operations and Gaussian filtering. Specific studies added further steps, such as artifact removal and hole-filling for optimization.
Feature extraction in dermatological diagnosis
Importance of features
In ML, feature extraction and selection are pivotal. Many studies on skin lesion diagnosis utilize a myriad of features, often rooted in the ABCD rule of dermatology, which covers asymmetry, border irregularity, color, diameter, and texture.
Some researchers also harness the power of deep neural networks, particularly CNNs, for feature extraction.
ABCD rule and its modern utility
The ABCD rule, established in 1985, offers criteria for early melanoma detection, focusing on asymmetry, border irregularity, color variation, and diameter over 6mm. This rule remains influential in contemporary research.
Detailed feature extraction techniques
Various methods have been employed to assess asymmetry in lesions, from calculating major and minor axes to measuring lesion solidity and variance. Similarly, to evaluate lesion borders, techniques have ranged from calculating compactness, solidity, and convexity to measuring perimeter errors and utilizing convex hulls.
Skin lesion classification with ML
Skin lesion classification, a crucial step in detecting skin cancer, relies on computer-aided systems that preprocess, segment, and extract features from images, culminating in lesion classification into various classes. The process uses extracted descriptors to provide information about PSLs.
ML models for CNNs due to their accuracy in image classification and feature extraction capabilities. Furthermore, some researchers have adopted pre-trained CNNs, while others have utilized ML methods like support vector machines (SVMs) and K-nearest neighbors (KNN) for their classification tasks.
Classification overview
This section encapsulates the results derived from the classification task of the reviewed articles. Some articles lacked a classification section, so no results are cited from these.
It is pivotal to note that for an equitable comparison of classification performance across multiple works, the task should be executed on identical datasets. However, most articles used diverse image sets, so results are presented for papers with comparable datasets.
Challenges and observations
Accuracy as a metric should be applied cautiously, particularly when imbalanced datasets. Some datasets, although yielding excellent results, had limitations in their size, questioning the trustworthiness of the outcome.
Also, innovative approaches, like the effects of darkening skin areas or balanced learning mechanisms, highlight the varied techniques employed in the field.
Australian research unlocks new path for cancer immunotherapy